Logan Mondal Bhamidipaty (罗根)

I'm a Stanford student pursuing an MS in CS focusing on AI and a BS in math. I'm fortunate to have worked with many amazing collaborators at the Stanford AI Lab (SAIL) including Emma Brunskill, Mykel Kochenderfer, Trevor Hastie, and Chelsea Finn. I also work part-time as an economic consultant with Paul Milgrom at Auctionomics.

As a researcher, I am broadly interested in the confluence of RL, game theory, and market design. In particular, I want to build RL tools that allow autonomous agents to behave strategically in complex, real-world markets.

Email  /  GitHub  /  CV  /  Google Scholar  /  Desmos

Research

ExpFamilyPCA.jl: A Julia Package for Exponential Family Principal Component Analysis

Exponential family principal component analysis (EPCA) outperforms traditional PCA on data binary, count, and compositional data (i.e., whenever data is from the exponential family and not simply Gaussian). ExpFamilyPCA.jl is the first package for EPCA in Julia and the first in any language to support EPCA for most common exponential family distributions.

CompressedBeliefMDPs.jl: A Julia Package for Solving Large POMDPs with Belief Compression

Belief compression is a technique for planning in large environments with state and outcome uncertainty. CompressedBeliefMDPs.jl, part of the popular POMDPs.jl community, provides a general framework for applying belief compression in large POMDPs and offers a suite of algorithms for sampling, compressing, and solving.

Learning to Explore in POMDPs with Informational Rewards

Standard exploration methods typically rely on random coverage of the state space or coverage-promoting exploration bonuses. However, in partially observed settings, the biggest exploration challenge is often posed by the need to discover information-gathering strategies – e.g., an agent that has to navigate to a location in traffic might learn to first check traffic conditions and then choose a route. In this work, we design a POMDP agent that gathers information about the hidden state, using ideas from the meta-exploration literature.

Scaling Contrastive Preference Learning for Vision Language Models (WIP)

Contrastive preference learning (CPL) is an extension of direct preference optimization (DPO) for arbitrary MDPs that has achieved promising results at small scales. Our project shows that CPL scales elegantly to high-dimensional, multimodal architectures and provides a simple, efficient method for aligning and steering autonomous agents.

DynaDojo: An Extensible Platform for Benchmarking Scaling in Dynamical System Identification

DynaDojo is a novel Python platform for developing and benchmarking data-driven dynamical systems identification algorithms. It prioritizes resource-efficient parallelization strategies for running on clusters and provides 7 baseline algorithms, 20 dynamical systems, and 3 benchmarking challenges off the shelf, with users easily able to add more.

Teaching

Algorithmic Game Theory Course Reader

The reader focuses on various topics across algorithmic game theory including auction and contest design, equilibrium analysis, cryptocurrencies, design of networks and network protocols, reputation systems, social choice, and social network analysis.

Market Design Course Reader

The reader focuses on three areas: (1) the design of matching algorithms to solve assignment problems, with applications to school choice, housing markets, and kidney exchanges; (2) the design of auctions to solve general resource allocation problems, with applications to the sale of natural resources, financial assets, radio spectrum, and advertising; and (3) the design of platforms and exchanges, with applications to internet markets. It emphasizes connecting economic theory to practical applications.

Last updated October 2024.